The present technology relates to an image processing apparatus, an image processing method, and a program, and more particularly, to an image processing apparatus, an image processing method, and a program, by which it is possible to suppress erroneous detection in the detection of a non-moving object, that is, in the detection of an object being left behind or of an object being carried away.
For example, as technology of detecting that a suspicious object has been left behind or exhibits shown in a showroom have been carried away, there is a method of detecting a part, which is not a background and does not move, as a non-moving object in an image.
A non-moving object detection method of detecting a non-moving object includes a method in which a stationary area is detected from an image and the newly detected stationary area is registered as a background to create a plurality of backgrounds, and even when a stationary object appears in front of another stationary object and the two stationary objects overlap each other, it is possible to separately detect the two stationary objects (for example, refer to WO/2009/005141).
Furthermore, a non-moving object detection method includes a method in which update and the like of a background are performed in a plurality of time scales to guarantee robustness against disturbance (for example, refer to Japanese Unexamined Patent Application Publication No. 2008-257693).
In addition, in the non-moving object detection method disclosed in WO/2009/005141, shaking and the like of trees may be erroneously detected as a stationary object (a non-moving object). Moreover, in the non-moving object detection method disclosed in WO/2009/005141, in order to separate stationary objects from each other, since various types of feature amounts are calculated and clustering is further necessary, an operation amount is increased, so that it is not suitable for a realtime process.
Furthermore, in the non-moving object detection method disclosed in Japanese Unexamined Patent Application Publication No. 2008-257693, when two stationary objects overlap each other, the two stationary objects are not guaranteed to be appropriately detected as immobile bodies.
Furthermore, according to representative detection technology using a background image in moving object detection, non-moving object detection and the like, a normal background may not be generated, and an erroneous result may be obtained in an area through which a person or an object such as a vehicle frequently passes. In order to cope with such a case, Japanese Unexamined Patent Application Publication No. 2004-265291 has proposed technology of calculating dispersion of a luminance value as a variation of each pixel to determine whether the area is an effective area which can be processed. Furthermore, a background difference method has been proposed to perform processes by excluding an area determined as an ineffective area.
However, according to the method proposed in Japanese Unexamined Patent Application Publication No. 2004-265291, since an effective area is determined based on standard deviation (dispersion) of a luminance value, it is not possible to discriminate only a situation in which a background is actually difficult to create. For example, when a background is detected using a multinomial distribution, the background is difficult to create in a situation in which a person frequently passes through, but the background can be generated in a situation in which the difference in luminance values is large such as flicker of a light. According to the method proposed in Japanese Unexamined Patent Application Publication No. 2004-265291, since the effective area is determined based on the dispersion of the luminance value, this may be determined as the same “ineffective area.” Therefore, since it is not possible to determine the presence or absence of an undetectable area due to the difficulty of background creation, accurate generation of a background may not be possible.
Furthermore, Japanese Unexamined Patent Application Publication No. 2006-331306 has proposed a method of determining a timewise change in each pixel as an intermediate class, an unstable class and the like. However, according to this determination method, since stability is determined based only on weights of a plurality of distributions, it is difficult to determine an area in which a change has occurred in a short time, or an actually difficult area due to frequent passage. Therefore, also in this case, accurate generation of a background may not be possible.